On 12/5/21 2:07 AM, Hannes Reinecke wrote:

On 12/4/21 9:13 PM, Jens Axboe wrote:

quoted

On 12/4/21 3:43 AM, Hannes Reinecke wrote:

quoted

On 12/3/21 10:45 PM, Jens Axboe wrote:

quoted

If we have a list of requests in our plug list, send it to the driver in
one go, if possible. The driver must set mq_ops->queue_rqs() to support
this, if not the usual one-by-one path is used.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
   block/blk-mq.c         | 24 +++++++++++++++++++++---
   include/linux/blk-mq.h |  8 ++++++++
   2 files changed, 29 insertions(+), 3 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 22ec21aa0c22..9ac9174a2ba4 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -2513,6 +2513,7 @@ void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
   {
   	struct blk_mq_hw_ctx *this_hctx;
   	struct blk_mq_ctx *this_ctx;
+	struct request *rq;
   	unsigned int depth;
   	LIST_HEAD(list);
   

@@ -2521,7 +2522,26 @@ void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
   	plug->rq_count = 0;
   
   	if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
-		blk_mq_run_dispatch_ops(plug->mq_list->q,
+		struct request_queue *q;
+
+		rq = plug->mq_list;
+		q = rq->q;
+
+		/*
+		 * Peek first request and see if we have a ->queue_rqs() hook.
+		 * If we do, we can dispatch the whole plug list in one go. We
+		 * already know at this point that all requests belong to the
+		 * same queue, caller must ensure that's the case.
+		 */
+		if (q->mq_ops->queue_rqs &&
+		    !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {

What is the dependency on shared tags here?
  From what I've seen it's just about submitting requests; the only
difference to shared tags is the way the tags are allocated.
Care to explain?

For shared tags, we need to actively increment the use count per
request. This path doesn't do that, so it's disabled for now. It could
be done, but then it'd have to be in the caller, so I'd rather leave it
for a future optimization if anyone cares enough about this for shared
tags.

I can add a comment about it if that helps.

Please do.
It'll act as a reminder what needs to be done if and when one of the 
drivers requiring shared tags is looking at implementing queue_rqs.

I added to the comment yesterday:

https://git.kernel.dk/cgit/linux-block/commit/?h=perf-wip&id=f9f526700607bf804fa8541c824ea54253f4241a

-- 
Jens Axboe